Abstract
Workflow mining is the task of automatically producing a workflow model from a set of event logs recording sequences of workflow events; each sequence corresponds to a use case or workflow instance. Formal approaches to workflow mining assume that the event log is complete (contains enough information to infer the workflow) which is often not the case. We present a learning approach that relaxes this assumption: if the event log is incomplete, our learning algorithm automatically derives queries about the executability of some event sequences. If a teacher answers these queries, the algorithm is guaranteed to terminate with a correct model. We provide matching upper and lower bounds on the number of queries required by the algorithm, and report on the application of an implementation to some examples.